AI agent vs hiring an operations assistant
An AI agent and an operations hire are not competing answers to the same question. An agent runs defined, repeating work consistently and escalates anything outside the definition. A person handles ambiguity, exceptions and relationships, and invents process where none exists. The deciding test is simple: if the job can be written down well enough that a new hire could follow it from the page, an agent can run it. If it cannot, you need the person.
The comparison most teams make is the wrong one
The usual framing is cost per month, agent against salary. It produces a clean looking number and a bad decision, because it assumes the two cover the same ground. They do not. Comparing them on price is like comparing a dishwasher to a chef on the grounds that both are involved with dinner.
The useful comparison is coverage. Write out the actual jobs sitting in the role you are trying to fill. Then sort them into two piles: the ones you could describe completely on a single page, and the ones where the description would end in the phrase use your judgment. The first pile is agent work. The second pile is why you hire.
Where the agent genuinely wins
Consistency on repetitive work, and it is not close. The report that gets built on the fifteenth of every month gets built on the fifteenth of every month, at the same quality, in December and in July. The list gets cleaned the same way the hundredth time as the first. The inbox gets sorted on the Friday before a holiday weekend.
This is not an intelligence advantage. It is an attention advantage. Repetitive work degrades when a person does it, not because people are careless but because the work is boring and something more urgent always appears. An agent has no competing priorities and no boredom, which makes it better than a person at precisely the jobs people least want.
Where the hire genuinely wins
Everything that starts with an ambiguous request. A client writes something slightly off and a good assistant hears the actual concern underneath it. A deadline slips and a person renegotiates it before it becomes a problem. A process does not exist yet and someone has to invent one, try it, and notice it is not working.
Agents escalate in these situations, which is correct behavior and is also not a solution. Escalation moves the decision to a human. If you have no human with time to receive it, the escalation sits.
The failure modes are different and both need managing
A person fails occasionally, visibly, and usually says so. An agent fails silently and identically on every run until someone checks, because it has no sense that today's output is worse than last week's. That is the single most important operational difference and it is why approval gates on consequential actions and a readable record of every run are structural requirements rather than nice features.
Judged against an unsupervised agent, a person is far safer. Judged against a supervised one, the comparison turns back into a question of scope.
The answer most teams land on
Both, in sequence. Hand the written down work to an agent first, because that work is currently slipping and nobody enjoys it. Then look at what is left. Often what is left is a genuinely interesting role that a good person will take and keep, rather than a role that is sixty percent drudgery and quietly churns every eighteen months.
Steward prices by skill count for this reason: a skill is one scoped, named job, so the plan is the scope. One skill is 500 dollars per month, up to three is 1500, up to five is 2500, with LLM tokens and server cost passed through.
Related: managed agents on a small team, what a managed AI agent is, and a worked example on inbox triage. More on the blog; about Steward.
Frequently asked questions
Is an AI agent cheaper than hiring an operations assistant?
Usually yes on direct cost, but that is the least useful way to compare them. An agent is priced per scoped job and a person is priced per unit of time and judgment. The honest comparison is what each one covers: an agent runs defined, repeating work indefinitely without supervision drift, and a person handles the undefined work, the exceptions and the relationships. Teams that replace a person with an agent for judgment work are usually unhappy within a quarter.
What can an operations assistant do that an AI agent cannot?
Everything that requires reading a room, absorbing an ambiguous request and deciding what the right outcome actually is. A person notices that a client is annoyed before the client says so, renegotiates a deadline, and invents a process where none exists. An agent executes a defined process reliably and escalates when the input does not match the definition. The dividing line is whether the job can be written down.
What can an AI agent do that an operations assistant cannot?
Run the same job at the same quality at 3am on a holiday, indefinitely, without the task quietly degrading because it is boring. Consistency on repetitive work is where agents outperform people, not because they are smarter but because they do not get bored, distracted or reassigned to something more urgent.
Can an AI agent replace an operations assistant entirely?
Only if that role is entirely defined, repeating work, which is rare and usually a sign the role was scoped too narrowly to begin with. The more common and more useful outcome is that the agent absorbs the repeating portion of the role so the person spends their time on exceptions, relationships and improvement. Teams generally find that raises what they get from the hire rather than removing the need for one.
How long does each take to get productive?
An agent is productive on a scoped job as soon as the job is defined and access is granted, which is typically days, and it is useless on anything outside that definition. A new hire is useful on ambiguous work almost immediately and takes weeks or months to reach full productivity on your specific systems and context. Fast on narrow versus slow on broad is the real trade.
What happens when each one fails?
They fail differently and both failure modes need managing. A person fails occasionally and visibly, and usually tells you. An agent fails silently and identically every run until someone checks, which is exactly why approval gates on consequential actions and a readable run record are not optional. An agent without those is an unsupervised employee who never reports a problem.